Re: [PATCH V2 1/3] lib: introduce some memory copy macros and functions

From: Miao Xie
Date: Thu Sep 02 2010 - 06:12:29 EST


On Thu, 02 Sep 2010 10:55:58 +0200, Andi Kleen wrote:
Miao Xie<miaox@xxxxxxxxxxxxxx> writes:

Changes from V1 to V2:
- change the version of GPL from version 2.1 to version 2

the kernel's memcpy and memmove is very inefficient. But the glibc version is
quite fast, in some cases it is 10 times faster than the kernel version. So I


Can you elaborate on which CPUs and with what workloads you measured that?

I did this test on x86_64 box with 4 cores, and the workload is quite low,
and I just do 500 bytes copy for 5,000,000 times.

the attached file is my test program.

The kernel memcpy is optimized for copies smaller than a page size
for example (kernel very rarely does anything on larger than 4k),
the glibc isn't. etc. There are various other differences.

memcpy and memmove are very different. AFAIK noone has tried
to optimize memmove() before because traditionally it wasn't
used for anything performance critical in the kernel. Has that
that changed? memcpy on the other hand while not perfect
is actually quite optimized for typical workloads.

Yes,the performance of memcpy on the most architecture is well,

But some of memmoves are implemented by byte copy, it is quite inefficient.
Unfortunately those memmove are used to modify the metadata of some filesystems,
such as: btrfs. That is memmove is importent for the performance of those filesystems.

So I improve the generic version of memcpy and memmove, and x86_64's memmove
which are implemented by byte copy.

One big difference between the kernel and glibc is that kernel
is often cache cold, so you e.g. the cost of a very large code footprint
memcpy/memset is harder to amortize.

Microbenchmarks often leave out that crucial variable.

I have some systemtap scripts to measure size/alignment distributions of
copies on a kernel, if you have a particular workload you're interested
in those could be tried.

Good! Could you give me these script?

Just copying the glibc bloat uncritical is very likely
the wrong move at least.

Agree!

Thanks!
Miao
#include <linux/module.h>
#include <linux/kthread.h>
#include <linux/time.h>
#include <linux/sched.h>
#include <linux/err.h>
#include <linux/string.h>
#include <linux/slab.h>

void get_start_time(struct timeval *tv)
{
do_gettimeofday(tv);
}

void account_time(struct timeval *stv, struct timeval *etv, int loops)
{
do_gettimeofday(etv);

if (loops) {
while (etv->tv_usec < stv->tv_usec) {
etv->tv_sec--;
etv->tv_usec += 1000000;
}

etv->tv_usec -= stv->tv_usec;
etv->tv_sec -= stv->tv_sec;

while (etv->tv_usec > 1000000) {
etv->tv_usec -= 1000000;
etv->tv_sec++;
}

printk("\tTotal loops: %d\n", loops);
printk("\tTotal time: %lds%ldus\n", etv->tv_sec, etv->tv_usec);
} else
printk("Didn't do any loop!\n");

}

char *str;

int init_module(void)
{
struct timeval stv, etv;
int loops, i;

str = kmalloc(1000, GFP_KERNEL);
if (!str)
return 0;
loops = i = 5000000;

printk("memcpy:\n");
get_start_time(&stv);
while (i--)
memcpy(str + 400, str, 500);
account_time(&stv, &etv, loops);

i = loops;
printk("\nmemmove:\n");
get_start_time(&stv);
while (i--)
memmove(str + 400, str, 500);
account_time(&stv, &etv, loops);

return 0;
}

void cleanup_module(void)
{
kfree(str);
}

MODULE_LICENSE("GPL");